UK Computer Scientists Are Trying to Build Real-Life Ultron - A Computer That Can Upgrade Itself
In the upcoming Avengers: Age of Ultron, Iron Man builds an AI that features humanlike intelligence, self-awareness and, most significantly, an ability to teach and upgrade itself. While the film takes the ultimate worst-case scenario stance on the topic (Ultron uses his powers of deduction to decide that humans are a blight on the Earth and should be exterminated), many artificial intelligence experts are terrified of a superintelligence singularity and what it would mean for humanity. On that note, computer scientists from DeepMind Technologies, a London-based artificial intelligence firm owned by Google, are trying to build a computer that, like Ultron, can program and upgrade itself.
The answer to creating an AI that has advanced learning capabilities, according to DeepMind, is bridging the gap between cutting-edge AI technology that imitates human neural networks and more basic computing technology. Neural technology can achieve astounding feats like facial recognition as a result of its proficiency in processing data input and classifying it into different categories. While basic computing technology is not as advanced, it is more proficient at rudimentary tasks, such as copying and storing data.
A machine that combines these two skill sets, which DeepMind calls the Neural Turing Machine (NTM), would be extremely formidable, as it would theoretically have the ability to process as much information as a computer while also possessing the learning and adapting capabilities of a human brain. In order to achieve this, the DeepMind computer scientists will essentially create a neural network, which will mimic the human brain, that also has a large external memory, which can store all the necessary data.
A preliminary study showed that the first iteration of NTM was able to copy binary code and sort lists of data with much greater speed and accuracy than other neural networks. Most intriguingly, the NTM learned much faster than other computing systems and, according to the researchers, used methods to complete its tasks that were very similar to a human programmer's. This is still a far cry from being able to program itself like Ultron, as the tasks it performed were very basic, but the results seem to confirm that DeepMind's methods increase a computer's ability to both learn and process data.
Other computer scientists agree that DeepMind's methods are ideal for creating an artificial intelligence that behaves like a human mind; Chris Eliasmith at the University of Waterloo said, "As humans we classify but we also manipulate the classification," he says. "If you want to build a computer that is cognitive in the way that we are, it is going to require this kind of control."